Amazon Bedrock¶

Overview¶

Fully managed service for foundation models (FMs).

Key Concepts¶

Concept	Description
Foundation Model	Large pre-trained model
Fine-tuning	Customize FM with your data
Guardrails	Content filtering and safety
Knowledge Bases	RAG with your data
Agents	Automated task execution

Available Models¶

Provider	Models
Amazon	Titan Text, Titan Embeddings, Titan Image
Anthropic	Claude 3 family
Meta	Llama 3 family
Cohere	Command, Embed
Stability AI	Stable Diffusion
Mistral	Mistral, Mixtral

Using Bedrock¶

Text Generation¶

import boto3
import json

bedrock = boto3.client("bedrock-runtime")

response = bedrock.invoke_model(
    modelId="anthropic.claude-3-sonnet-20240229-v1:0",
    contentType="application/json",
    accept="application/json",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 1000,
        "messages": [
            {"role": "user", "content": "Explain machine learning in simple terms."}
        ]
    })
)

result = json.loads(response["body"].read())

Embeddings¶

response = bedrock.invoke_model(
    modelId="amazon.titan-embed-text-v1",
    contentType="application/json",
    accept="application/json",
    body=json.dumps({
        "inputText": "Sample text for embedding"
    })
)

Fine-tuning¶

Customize models with your data.

bedrock = boto3.client("bedrock")

response = bedrock.create_model_customization_job(
    jobName="my-fine-tuning-job",
    customModelName="my-custom-model",
    roleArn=role_arn,
    baseModelIdentifier="amazon.titan-text-express-v1",
    trainingDataConfig={
        "s3Uri": "s3://bucket/training-data/"
    },
    outputDataConfig={
        "s3Uri": "s3://bucket/output/"
    },
    hyperParameters={
        "epochCount": "3",
        "batchSize": "8",
        "learningRate": "0.00001"
    }
)

Guardrails¶

Content moderation and filtering.

Denied topics
Content filters (hate, violence, etc.)
Word filters
PII detection and masking

Knowledge Bases (RAG)¶

Augment models with your data.

graph LR
    A[User Query] --> B[Embedding]
    B --> C[Vector Search]
    C --> D[Retrieved Context]
    D --> E[FM + Context]
    E --> F[Response]

Agents¶

Automate multi-step tasks.

Define actions via Lambda functions
Access knowledge bases
Orchestrate complex workflows

SageMaker vs Bedrock¶

Feature	SageMaker	Bedrock
Custom Training	Yes	Limited (fine-tuning)
Foundation Models	Via JumpStart	Native
Infrastructure	Managed	Serverless
Use Case	Custom ML	GenAI applications

Exam Tips¶

!!! warning "Key Points" - Bedrock is serverless, no infrastructure management - Use fine-tuning for domain-specific customization - Guardrails for content safety - Knowledge Bases for RAG patterns - Agents for complex task automation